Hierarchical Causal Learning for Face Age Synthesis

This project is the official implementation of the paper "Hierarchical Causal Learning for Face Age Synthesis".

We propose a novel Hierarchical Causal Face Synthesis (HCFace) framework that automatically discovers causal relationships among facial attributes and leverages them to guide age synthesis.

Key Features

• Causal Graph Discovery Module
  Automatically models the causal relationships of facial attributes, discovers the causal structure among attributes, and constructs hierarchical causal graphs to guide subsequent age editing.

• Non-linear Mapping Module
  Takes the discovered hierarchical causal graph as input and guides the model to modify attribute values along the causal paths, generating facial images that realistically reflect facial aging patterns across different age groups.

Datasets

Training Dataset

We recommend train model using the MS1M or CASIA-Webface dataset, a large-scale face recognition dataset.

Dataset	Description	Download Link
MS1M (MS-Celeb-1M)	~10M images, 100k identities. We use the cleaned faces_emore version.	Link
CASIA-Webface	10K ids/0.5M images	Link

Testing Datasets

We evaluate our method on four public age synthesis / age progression benchmarks:

Dataset	Description	#Images	#Subjects	Download Link
CACD	Cross-Age Celebrity Dataset	163,446	2,000	Link
FG-NET	Face Aging Dataset	1,002	82	Link
MORPH2	Longitudinal Face Dataset	55,134	13,618	Link
ECAF	dataset with a diverse age distribution, comprising 5,265 face images of 613 individuals, with an average age of 41.3 years.	-	-	Link

Setup

Environment Requirements

• Python 3.8+
• PyTorch 1.8+
• CUDA 11.1+ (recommended)

Installation

1. Clone this repository:
2. git clone https://github.com/SE-hash/HCFace.git
3. cd HCFace
4. pip install requirement.txt
5. Follow main.py, replace your GPU number and dataset name. Note that you should ensure that the number of compute cards you can use matches the number specified in --nproc_per_node=x